Learning singing from speech

Abstract

A method, computer program, and computer system is provided for converting a singing voice of a first person associated with a first speaker to a singing voice of a second person using a speaking voice of the second person associated with a second speaker. A context associated with one or more phonemes corresponding to the singing voice of a first person is encoded, and the one or more phonemes are aligned to one or more target acoustic frames based on the encoded context. One or more mel-spectrogram features are recursively generated from the aligned phonemes, the target acoustic frames, and a sample of the speaking voice of the second person. A sample corresponding to the singing voice of a first person is converted to a sample corresponding to the second singing voice using the generated mel-spectrogram features.

What is claimed is: 
     
       1. A method of converting a singing voice of a first person to a singing voice of a second person using a speaking voice of the second person, comprising:
 encoding, by a computer, a context associated with one or more phonemes corresponding to the singing voice of a first person; 
 aligning, by the computer, the one or more phonemes to one or more target acoustic frames based on the encoded context; 
 subsequent to the aligning the one or more phonemes to the one or more target acoustic frames, recursively generating, by the computer, one or more mel-spectrogram features from the aligned phonemes, the target acoustic frames, and a sample of the speaking voice of the second person based on one or more frame-aligned hidden states being used as input for autoregressive generation, wherein the recursively generating the one or more mel-spectrogram features comprises performing a post-CBHG technique to an output from each recursion; and 
 converting, by the computer, a sample corresponding to the singing voice of a first person to a sample corresponding to the second singing voice using the generated mel-spectrogram features. 
 
     
     
       2. The method of  claim 1 , wherein the encoding comprises:
 receiving a sequence of the one or more phonemes; and 
 outputting a sequence of one or more hidden states containing a sequential representation associated with the received sequence of phonemes. 
 
     
     
       3. The method of  claim 2 , wherein the aligning the one or more phonemes to one or more target acoustic frames comprises:
 concatenating the output sequence of hidden states with information corresponding to the singing voice of a first person; 
 applying dimension reduction to the concatenated output sequence using a fully connected layer; 
 expanding the dimension-reduced output sequence based on a duration associated with each phoneme; and 
 aligning the expanded output sequence to the target acoustic frames. 
 
     
     
       4. The method of  claim 3 , further comprising concatenating one or more frame-aligned hidden states with a frame level, a root mean square error value, and a relative position associated with every frame. 
     
     
       5. The method of  claim 4 , wherein the duration of each phoneme is obtained from a force alignment performed on one or more input phonemes and one or more acoustic features. 
     
     
       6. The method of  claim 1 , wherein the generating the one or more mel-spectrogram features based on the aligned frames comprises:
 computing an attention context from one or more encoded hidden states aligned with the one or more target acoustic frames; and 
 applying a CBHG technique to the computed attention context. 
 
     
     
       7. The method of  claim 6 , wherein a loss value associated with the mel-spectrogram is minimized. 
     
     
       8. The method of  claim 1 , wherein the generating the one or more mel-spectrogram features is performed by a recursive neural network. 
     
     
       9. The method of  claim 8 , wherein the inputs to the recursive neural network comprise a sequence of the one or more phonemes, a duration associated with each of the one or more phonemes, a fundamental frequency, a root mean square error value, and an identity associated with a speaker. 
     
     
       10. The method of  claim 1 , wherein the singing voice of a first person is converted to the second singing voice without parallel data and without changing the content associated with the singing voice of a first person. 
     
     
       11. A computer system for converting a singing voice of a first person to a singing voice of a second person using a speaking voice of the second person, the computer system comprising:
 one or more computer-readable non-transitory storage media configured to store computer program code; and 
 one or more computer processors configured to access said computer program code and operate as instructed by said computer program code, said computer program code including:
 encoding code configured to cause the one or more computer processors to encode a context associated with one or more phonemes corresponding to the singing voice of a first person; 
 aligning code configured to cause the one or more computer processors to align the one or more phonemes to one or more target acoustic frames based on the encoded context; 
 generating code configured to cause the one or more computer processors to, subsequent to the aligning the one or more phonemes to the one or more target acoustic frames, recursively generate one or more mel-spectrogram features from the aligned phonemes, the target acoustic frames, and a sample of the speaking voice of the second person based on one or more frame-aligned hidden states being used as input for autoregressive generation, wherein the recursively generating the one or more mel-spectrogram features comprises performing a post-CBHG technique to an output from each recursion; and 
 converting code configured to cause the one or more computer processors to convert a sample corresponding to the singing voice of a first person to a sample corresponding to the second singing voice using the generated mel-spectrogram features. 
 
 
     
     
       12. The system of  claim 11 , wherein the encoding code is configured to cause the one or more processors to:
 receive a sequence of the one or more phonemes; and 
 output a sequence of one or more hidden states containing a sequential representation associated with the received sequence of phonemes. 
 
     
     
       13. The system of  claim 12 , wherein the aligning code is configured to cause the one or more processors to:
 concatenate the output sequence of hidden states with information corresponding to the singing voice of a first person; 
 apply dimension reduction to the concatenated output sequence using a fully connected layer; 
 expand the dimension-reduced output sequence based on a duration associated with each phoneme; and 
 align the expanded output sequence to the target acoustic frames. 
 
     
     
       14. The system of  claim 13 , further comprising concatenating code configured to cause the one or more processors to concatenate one or more frame-aligned hidden states with a frame level, a root mean square error value, and a relative position associated with every frame. 
     
     
       15. The system of  claim 14 , wherein the duration of each phoneme is obtained from a force alignment performed on one or more input phonemes and one or more acoustic features. 
     
     
       16. The system of  claim 11 , wherein the generating the one or more mel-spectrogram features based on the aligned frames comprises:
 computing an attention context from one or more encoded hidden states aligned with the one or more target acoustic frames; and 
 applying a CBHG technique to the computed attention context. 
 
     
     
       17. The system of  claim 11 , wherein the generating the one or more mel-spectrogram features is performed by a recursive neural network. 
     
     
       18. The system of  claim 17 , wherein the inputs to the recursive neural network comprise a sequence of the one or more phonemes, a duration associated with each of the one or more phonemes, a fundamental frequency, a root mean square error value, and an identity associated with a speaker. 
     
     
       19. The system of  claim 11 , wherein the singing voice of a first person is converted to the second singing voice without parallel data and without changing the content associated with the singing voice of a first person. 
     
     
       20. A non-transitory computer readable medium having stored thereon a computer program for converting a singing voice of a first person to a singing voice of a second person using a speaking voice of the second person, the computer program configured to cause one or more computer processors to:
 encode a context associated with one or more phonemes corresponding to the singing voice of a first person; 
 align the one or more phonemes to one or more target acoustic frames based on the encoded context; 
 subsequent to the aligning the one or more phonemes to the one or more target acoustic frames, recursively generate one or more mel-spectrogram features from the aligned phonemes, the target acoustic frames, and a sample of the speaking voice of the second person based on one or more frame-aligned hidden states being used as input for autoregressive generation, wherein the recursively generating the one or more mel-spectrogram features comprises performing a post-CBHG technique to an output from each recursion; and 
 convert a sample corresponding to the singing voice of a first person to a sample corresponding to the second singing voice using the generated mel-spectrogram features.